How Bloggers Scam You With Fake Claude Watermarks and GitHub Scripts

How Bloggers Fool You with “Secret Claude Prompts”

And why “secret prompts” for AI text humanization are just a traffic scam and a malware risk for your PC

Picture this: you are scrolling through your feed and a blogger’s hyped-up video pops up. “Urgent! How to bypass invisible Claude watermarks in one click! Just follow the GitHub link, download the file, paste it into the AI, and your text becomes 100% human!”

РYour hand just reaches out to click. After all, nobody wants a university professor or a strict editor to tear you apart for a “machine style.” Stop. Let us exhale, turn on critical thinking, and ruthlessly expose this scheme. First, because there are simply no invisible watermarks in Claude. Second, because downloading such files from GitHub could cost you every password to your corporate accounts (just ask the former Disney employee). Let us break down how they try to scam us for reach and expose us to a real cyber threat.

Pop art illustration exposing online PTC scams, fake get-rich-quick methods, and ad-clicking myths.
Discover the truth behind NeoBux and PTC scams. Learn why clicking ads for $20 a day is a mathematical trap

The core of the offer: how they trick you into boosting bloggers’ reach 

The blogger talks about how to humanize text. Everything is simple:

  1. Follow the link to GitHub;
  2. Download an unknown file from an unknown source to your computer or phone (hello, phishing);
  3. Open the Skills section in Claude, and add this file there;
  4. Run your text through this custom skill and copy the result;
  5. Then check the text through an anti-AI detector, which will show “100% human.”

What is the trick behind this absurdity?

  1. First, anti-AI detectors are extremely unstable in principle and produce false results. I have already written about this here. 
  2. Second, there are no watermarks at all. 
  3. Third, to get this “super prompt,” you are required to subscribe to the channel, leave a code word in the comments, or drop your email. In short, you boost the blogger’s reach. 

The scheme is as reliable as a Swiss watch: the generated text has no watermarks to begin with, you go to GitHub, download a prompt, paste it into a skill that supposedly removes non-existent marks, copy the text into a detector, and voila, “written by a human.” That is, the author simply fed the text back to the same Claude with a request to “remove AI markers,” naively believing that since they were supposedly there, they are now gone. This is like staging a whole show with regular soda: you cannot just pour it into a glass and hand it to a client. First you have to pour it into a shaker, shake it up, toss in some fancy ice, a little umbrella, a fruit slice, and a straw. Basically, the contents of the glass have not changed at all, but the presentation is spectacular!

What are “Skills” in Claude and why the scheme of cleaning something non-existent is total absurdity

In this context, a Skill refers to a mini-instruction or plugin that a user can upload directly into the neural network to train it to perform a specific task in a tailored way. 

Think of them as shortcut commands: you paste a pre-made prompt just once, and from then on, this skill works seamlessly in the background, delivering fast results without requiring you to re-paste the instruction before every single message. For comparison: in Gemini, these are called Gems, while in ChatGPT, they are Custom GPTs.

What Claude actually thinks about “invisible tags”

I do not intentionally insert any hidden or invisible characters into my responses, neither watermark characters, nor zero-width spaces, nor any special tags that would distinguish text as “AI-generated” at the character level. What you see in the chat is plain text, along with Markdown formatting if it is used for styling.
  • Technical nuances: theoretically, this is possible. There are studies and experimental approaches, such as Google’s SynthID technology for text, which embed statistical watermarks at the token selection level. These are not separate invisible characters, but a subtle shift in probabilities during word generation that can be detected statistically, but cannot be seen with the naked eye or found as a hidden symbol. 
  • Anthropic’s stance: as for Claude specifically, there is no publicly announced text watermarking feature. 
  • If you see strange characters when copying text into another editor, this is usually related to encoding, auto-replacement of quotes, dashes, or Markdown special characters, rather than “surveillance.”

Real threat: why “useful files from GitHub” can leave you jobless 

If you are not a developer, going to GitHub and downloading random executable files or scripts is a huge risk. 

In July 2024, there was an eye-opening case involving a Disney employee who decided to optimize his routine work using AI. Mid-level technical manager Matthew van Andel downloaded an unofficial free content generation tool from GitHub. 

The downloaded program contained hidden malware, specifically a data stealer and a keylogger. The virus quietly ran on his computer for about five months, gaining access to the employee’s password manager, where two-factor authentication was not enabled. Through the stolen credentials, attackers broke into Disney’s corporate Slack. 

As a result, more than 1 terabyte of internal data was compromised and leaked online, including about 44 million messages from work chats, financial reports, and personal employee data. All work files leaked into the public domain, and the employee himself was fired in disgrace.

The only working “anti-AI detector” is your own head

Claude does not put any physical watermarks anywhere: pure probability mathematics and word choice are at work here. The whole idea of using “hacker” schemes via GitHub and custom skills to clear out non-existent watermarks is strictly content created for hype, views, and monetizing the audience’s gullibility. 

The most sensible and safe workflow algorithm:

  • Generate a draft with the AI so it lays out the structure, facts, and core points.
  • Read through the text with your own eyes and critically evaluate the facts.
  • Humanize the text manually by removing overly smart, bookish clichés, adding dynamic flow, and tweaking a couple of sentences so they sound like your usual conversational style.

This is a thousand times more effective, faster, and safer than any sketchy “remover skills.” A lively authorial style and your own editing represent the best anti-plagiarism and the most reliable text humanizer.

Generally speaking, the level of usefulness from such “secret prompts” is about the same as a “Caution: Wet Floor” sign right in the middle of the Sahara Desert…

Insider Tip: to truly protect your workflow from hidden infostealers, you must understand that downloaded scripts from untrusted repositories are the primary vector for unauthorized credential harvesting. If you want to dive deeper into how modern malware bypasses antivirus software and how to secure your local machine against trojanized developer tools, I highly recommend reading The Art of Software Security Assessment by Mark Dowd et al. It is the ultimate manual on recognizing vulnerable code execution and securing your environment without falling for hype-driven downloads.

Anyway, while people search for nonexistent watermarks and risk downloading malware from GitHub, I’m focused on creating things that are actually safe, practical, and inspiring. if you want to upgrade your workspace, check out these useful digital items on Gumroad, or pick up a cute and cool design from my Redbubble shop! 🦝

Now confess: how many times have you downloaded “secret prompts” yourself in exchange for a subscription and a code word in the comments? 🤡

Support the Lab!

Liked this article?  If you found this article helpful, please share it with your friends and subscribe to my blog! I regularly share a high-quality mix of practical daily hacks, unusual real-life stories, and fascinating debunking articles. There is plenty of useful and interesting content coming your way. Subscribe to my newsletter, follow me on Medium and Substack .

Disclaimer: this article contains affiliate links. As an Amazon Associate, I earn from qualifying purchases. If you make a purchase through these links, I may earn a small commission at no extra cost to you.

Comments

Post a Comment